pseudocode.txt

for each epoch:
    Sample a batch of attacks (e.g. 8): PGD x 2, FSGM x 4, SIM x 2
    Initialize theta = []
    for each attack:
        Sample a batch of images (32) with 2 augmented views: 3 x (32 x 3 x 224 x 224)
        1. Get x_i: 32 x 3 x 224 x 224
        2. Get x_j: 32 x 3 x 224 x 224
        3. Use selected attack & current model parameter to generated x_adv: 32 x 3 x 224 x 224
        4. Feed into encoder & projector
        5. Compute loss
        6. Compute updated model parameter and append in theta
    Initialize loss = []
    for each theta0 in theta:
        1. load theta0 in model
        2. Sample another batch of images (32) with 2 augmented views: 3 x (32 x 3 x 224 x 224)
        3. Get x_i: 32 x 3 x 224 x 224
        4. Get x_j: 32 x 3 x 224 x 224
        5. Use selected attack & current model parameter to generated x_adv: 32 x 3 x 224 x 224
        6. Feed into encoder & projector
        7. Compute local loss
        8. Append local loss in loss
    Compute mean/sum of loss
    Meta-update global model parameters